Abstract
Generative flows are attractive because they admit exact likelihoodoptimization and efficient image synthesis. Recently, Kingma & Dhariwal (2018)demonstrated with Glow that generative flows are capable of generating highquality images. We generalize the 1 x 1 convolutions proposed in Glow toinvertible d x d convolutions, which are more flexible since they operate onboth channel and spatial axes. We propose two methods to produce invertibleconvolutions that have receptive fields identical to standard convolutions:Emerging convolutions are obtained by chaining specific autoregressiveconvolutions, and periodic convolutions are decoupled in the frequency domain.Our experiments show that the flexibility of d x d convolutions significantlyimproves the performance of generative flow models on galaxy images, CIFAR10and ImageNet.